Research On Patent Quality Evaluation Method Based On Knowledge Graph

Scientific and technological innovation and technological progress are the core competitiveness of the country,and innovation contributes to a high level of scientific and technological self-reliance and self-improvement.As one of the important products of technological innovation,patents contain the core elements of innovation,and their quality is valued by the country and the enterprise.The improvement of patent quality is of great significance to the country and the enterprise.Currently,automated patent quality evaluation enables efficient screening and analysis of patents,enhancing the accuracy and efficiency of evaluation,thus driving the improvement of patent quality.However,current automated methods for patent quality evaluation predominantly concentrate on analyzing evaluation indicators and textual data,while neglecting the analysis of intricate relationships among patents.Such complex relationships are manifested as multi-relationship and multi-attribute networks among patents,such as citation relationships,inventor relationships,and the number of independent claims among patents.Therefore,this thesis addresses the above deficiencies by firstly constructing a patent quality evaluation dataset to obtain the multi-relationship and multi-attribute data of patents.Secondly,the multi-relationship and multi-attribute network is modeled as a patent knowledge graph as the basis for obtaining the initial features of patents.Finally,a graph neural network method based on the knowledge graph is proposed to predict the patent quality.The main work of this thesis are as follows:(1)We constructed a patent quality evaluation dataset.Firstly,In this thesis,we selected patents in the field of artificial intelligence as the research objects.Secondly,we defined patent quality as its impact on subsequent patents and used this criterion to establish a classification standard for patent quality.Lastly,we collected data using a patent database platform and annotated patent samples according to the quality classification criteria,thus completing the construction of the patent quality evaluation dataset.(2)We constructed a patent knowledge graph.This focuses on the task of patent quality evaluation and adopts a top-down approach to construct a patent knowledge graph,which involves three main components: ontology modeling,information extraction,and knowledge storage.In the ontology modeling stage,the entity relationships and attributes related to patent quality are selected and instantiated by consulting experts in related fields to construct the patent knowledge graph ontology model.In the information extraction stage,the original patent data is converted into knowledge graph triad data by attribute triad extraction and relationship triad extraction.In the knowledge storage stage,the triad data are imported into the graph database for storage and visualization analysis.(3)We propose a method for patent quality evaluation of Knowledge Graph based Relational Graph Neural Network(KGRGNN).The proposed method first utilizes a knowledge representation learning model to obtain the embedding vectors of entity nodes and relationships in the patent knowledge graph.Subsequently,addressing the limitation of traditional graph neural networks in fully exploiting the semantic information contained in the multi-relational and multi-attribute patent knowledge graph,this study introduces a relational graph convolutional network.This network aggregates the representations of neighboring entities that encode relational semantic information,enabling KGRGNN to capture high-order neighborhood semantic information of patent entities across different relationships.Finally,extensive experiments are conducted on the patent quality evaluation dataset,and the results demonstrate the effectiveness of the KGRGNN method.